Occlusal Geometry in Closed Form for Orthodontic Report Generation
Ajo Babu George, Govind Arun, Sidharth N Krishna, Uma Ranjan
cs.CV · cs.CL
Quick verdict
Casts Bite2Text report generation as closed-form occlusal measurement plus templates; third in ODIN 2026, with a sharp critique of the recall-heavy evaluator.
Summary
The paper recasts the ODIN 2026 Bite2Text orthodontic report generation task as closed-form geometric measurement rather than multimodal captioning. Because the released scan pairs arrive registered in occlusion, the authors recover an anatomical frame per case (the stated RAS convention fails in the release), reduce each arch to an occlusal ridge profile in arch-angle coordinates, and derive 31 quantities — overbite, overjet, midline deviation, crossbite extent, the curves of Spee and Wilson, interdigitation lag — in closed form. Gradient boosting maps these onto 13 template fields, predicting a field only where patient-level cross-validation beats its majority baseline; fusing a ConvNeXt-Tiny classifier over five photograph views raises mean field accuracy from 0.601 to 0.683, and a deterministic template renders the six-part narrative. Reimplementing the evaluator shows its BLEU-4 and METEOR are local variants weighting recall nine-to-one, that two clinicians agree on 47% of findings for the same patient, and that a constant report outscores a genuine second clinician report by 0.165. Held out, the system reaches BLEU-4 0.458 and METEOR 0.677 against scan references and 0.278/0.507 against photograph references; it placed third in the challenge at 0.2680 BLEU-4, within 0.022 of first, on CPU in under ten seconds per case. Limitations: weakest sagittal classification (0.53–0.59), RadFact with a substitute judge, a photograph human anchor of seven paired cases, and laterality from 32 auxiliary cases.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Because the released Bite2Text scans are pre-registered in occlusion, occlusal quantities are measurable in closed form, and a measure-then-render system is competitive with multimodal LLM entries (third place at 0.2680 BLEU-4, within 0.022 of first, CPU-only under ten seconds per case).
- The deployed challenge evaluator uses local BLEU-4 and METEOR variants whose F-mean weights recall nine to one, under which a constant corpus-medoid report outscores a genuine second clinician report by 0.165 captioning.
- Two clinicians agree on only about 47 percent of findings for the same patient (finding Jaccard 0.469 over 498 pairs), bounding the achievable score against a single reference.
- Selective per-field prediction gated on beating each field's majority baseline, plus per-field fusion with a ConvNeXt-Tiny photograph classifier, raises mean field accuracy from 0.601 to 0.683.
- Geometry-derived reports cover more distinct entailed reference findings than the prior (4.29 vs 3.85 phrases) even though the prior scores marginally higher RadFact F1 (0.468 vs 0.459) under a substitute judge.
Strengths
Right architectural insight, grounded in dataset properties
Sections 1 and 3: pre-registered occlusal scan pairs make overbite, overjet, crossbite extent and the occlusal curves deterministic functions of the meshes, so a closed-form ridge profile in arch-angle coordinates (pole at ridge-extent midpoint, ±150 degrees, 1-degree bins) replaces perception, yielding 31 measurements with NumPy/SciPy/trimesh.
Evaluator forensics with quantitative teeth
Sections 2 and 4 with Fig. 1: a byte-faithful reimplementation shows the deployed METEOR variant weights recall 9:1 (F = 10PR/(R+9P); an unmatched extra token costs harmonic METEOR 55 percent but METEOR-lite only 16 percent), and Table 1 shows a corpus-medoid constant report (0.547 captioning) beating a genuine second clinician (0.382, n=145 scan pairs).
Measurement validation against an independent archive
Honest negative results and calibration discipline
Deployment realism
Weaknesses
Circular supervision
Metric gaming is partly embraced, not only critiqued
No significance testing on headline deltas
Thin human anchors under key claims
Fusion result unvalidated on the hidden test
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper addresses Task 2 (Bite2Text) of the ODIN 2026 challenge: generate a free-text orthodontic report from a registered upper/lower intraoral scan pair plus five standardized photographs. Its thesis is structural: because the released scans arrive mutually registered in occlusion, most quantities the report names (overbite, overjet, midline deviation, transverse overlap and crossbite, curves of Spee and Wilson, interdigitation lag, crowding) are deterministic functions of the two meshes, so the right architecture is measure-then-render, not multimodal perception. Three contributions are claimed and delivered: a byte-faithful reimplementation of the deployed evaluator exposing local BLEU-4/METEOR variants (Section 2); a 31-measurement closed-form geometry pipeline with conservative selective field prediction and photograph fusion (Section 3); and a calibrated score analysis showing the captioning metric rewards corpus central tendency over clinical accuracy (Section 4). The abstract's headline that a constant report outscores a genuine second clinician by 0.165 is backed by Table 1's second-clinician row (0.382 captioning on the scan family, n=145).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the Bite2Text task in the ODIN 2026 challenge?
Bite2Text is Task 2 of the ODIN 2026 challenge: given a registered pair of upper and lower intraoral scans plus five standardized intraoral photographs, produce a free-text orthodontic report covering transverse relations and crossbite, overbite, Angle sagittal classification, midline coincidence, the curves of Spee and Wilson, and crowding. Submissions were ranked by RadFact F1 at weight 0.8 plus the mean of BLEU-4 and METEOR at weight 0.2.
Why does a constant report beat a real clinician's report in this paper's evaluation?
The authors found the challenge's deployed METEOR variant weights recall nine times as heavily as precision, so omitting a topic is expensive while adding wrong assertions is nearly free. A constant corpus-medoid report therefore scores 0.547 captioning while a genuine second clinician's report of the same patient scores 0.382, because two clinicians agree on only about 47 percent of findings.
Did this system win the ODIN 2026 Bite2Text challenge?
No. It placed third on the hidden test at BLEU-4 0.2680 and METEOR 0.4629, within 0.022 BLEU-4 of the winning multimodal-LLM entry (0.2902), running on CPU in under ten seconds per case. The photograph-fusion improvement described in the paper was completed after the test phase and did not contribute to the ranked result.
Is the code and data for this paper available?
The paper states the dataset and code are available at github.com/GIND123/ODIN_toothfairy4. The pipeline is deterministic CPU-only geometry plus a roughly 5 MB gradient-boosting bundle, so held-out results are reproducible; hidden test-set scores cannot be reproduced because that set is unpublished.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The metadata abstract cuts off mid-sentence, but the full embedded text contains the complete abstract and body through the references; no true truncation was found at the end of the provided text.
- In Table 2, rows run together in the extraction (the oracle row's final value 0.699 is immediately followed by the next row label as '0.6990 4 6 dental health sentences asserted'); values were cross-checked against the surrounding prose, but one digit could be misread.
- Figure axis and tick labels (Fig. 3 field names, Fig. 4(b) submission names such as MIGG/MMLVM, bitemark final2, DiceMed) are interleaved with body text by the extractor; competitor submission names could be misparsed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.